by Brian Hatch
|Overview of LIDS, Part Three
last updated November 12, 2001
This is the third part of a four-part article devoted to the exploration of LIDS, a Linux kernel patch that will allow users to take away the all-powerful nature of root. The first article in this series offered an overview of LIDS. The second installment looked at file restrictions, LIDS File ACLs, and LIDS enhancements of Linux capabilities. This installment will discuss granting capabilities, the LIDS-specific capabilities, ACL inheritance and time-based ACLs.
Granting Capability Exceptions
In the traditional Linux model, a capability that is removed from the bounding set is not available to any process thereafter. LIDS, however, gives us a way to grant access to capabilities on a per program basis. Thus we could remove a capability from general usage and provide it only to the functions that need it.
As an example, let's take getty, which is used for logging in via the console or a serial port. It requires the CAP_SETUID and CAP_SETGID capabilities to become the user that logs in, and CAP_FOWNER, CAP_CHOWN, and CAP_DAC_OVERRIDE to muck with the tty and related file permissions.
If you removed these capabilities from the system by putting a '-' in front of them in lids.cap then getty would not be able to perform these actions (and neither would any other program). We will grant a capability (the Object) to a particular program (the Subject) using the lidsadm command. Assuming our getty binary is /sbin/getty, we'd run the following:
lfs# lidsadm -A -s /sbin/getty -o CAP_SETUID -j GRANT lfs# lidsadm -A -s /sbin/getty -o CAP_SETGID -j GRANT lfs# lidsadm -A -s /sbin/getty -o CAP_CHOWN -j GRANT lfs# lidsadm -A -s /sbin/getty -o CAP_FOWNER -j GRANT lfs# lidsadm -A -s /sbin/getty -o CAP_DAC_OVERRIDE -j GRANT
Getty will be given the named capabilities even though they are removed from the bounding set. As another example, say we want to be able to syncronize our clock using ntpdate, the following rules will grant the capabilities we need:
lfs# lidsadm -A -s /usr/sbin/ntpdate -o CAP_NET_BIND_SERVICE -j GRANT lfs# lidsadm -A -s /usr/sbin/ntpdate -o CAP_SYS_TIME -j GRANT
In this case CAP_NET_BIND_SERVICE lets us bind a low numbered port (123) and CAP_SYS_TIME lets us change the system clock.
While integrating itself into the Linux kernel, LIDS took the opportunity to add and enhance some capabilities to give us an extra layer of security and configurability.
Special Handling Of CAP_BIND_NET_SERVICE
Normally when a program has CAP_BIND_NET_SERVICE capability it is able to bind any port less than 1024. LIDS has extended this capability to allow us to define which port or ports a program can bind. To use the normal behaviour, you simply grant the service as follows:
lfs# lidsadm -A -s /usr/sbin/httpd -o CAP_BIND_NET_SERVICE -j GRANT
However, if you wanted to restrict Apache to bind only port 80 (HTTP), you'd use the following:
lfs# lidsadm -A -s /usr/sbin/httpd \ -o CAP_BIND_NET_SERVICE 80-80 -j GRANT
In order to allow it to bind port 443 (HTTPS) as well, you would use:
lfs# lidsadm -A -s /usr/sbin/httpd \ -o CAP_BIND_NET_SERVICE 80-80,443-443 -j GRANT
You can grant a range of ports using '-', and a list of ports by joining ranges with ','. For example, if we want to allow Stunnel to listen to IMAPS (993), IRCS (994), POP3S (995) and SSMTP (465) to provide SSL services, you'd have the following:
lfs# lidsadm -A -s /usr/sbin/stunnel \ -o CAP_BIND_NET_SERVICE 465-465,993-995 -j GRANT
This fine granular control allows you to grant programs the ability to bind only the ports they actually need, rather than using the sledgehammer approach.
The LIDS-specific capability CAP_INIT_KILL allows you to prevent daemons from being sent any signals. If you turn off this capability (by putting a '-' in front of the line in lids.cap) then no user or program will be able to signal a system daemon, where daemon is defined as a program who is a child of the init process (pid==1).
This is an excellent way to make sure that an attacker is not able to stop your log analysis tools, syslog daemon, cron, or other important services. The drawback is that you too cannot kill these programs. In fact, you can't send them any signals whatsoever, such as sending Apache a HUP to reread its config file after a legitimate change. To send any signal to daemons you will need to be in an LFS.
Also, some daemons send signals to their parent or children, Apache being an example, so you may need to grant these processes the CAP_INIT_KILL capability as follows:
lfs# lidsadm -A /usr/sbin/httpd \ -o CAP_INIT_KILL -j GRANT
I tend to like disabling CAP_INIT_KILL in lids.conf, but it does require that extra step to use an LFS in times of need.
CAP_HIDDEN is another LIDS-specific capability that is useful for hiding processes. Any program that is given the CAP_HIDDEN capability will not be visible in /proc (and thus not available to ps, top, and friends) at all.
This is not a 100% sure way of hiding processes, however. If you hide sshd, for example, folks will still be able to use 'netstat -na' and see that something is listening on port 22. A user could attempt to send a signal to each process ID from 1-65535 and determine that a hidden process existed when the signal was rejected or sent but no /proc entry exists for it. Or it could be discovered if it logs its PID to /var/log/NAME.pid.
However, I still find it useful, and like to hide some system processes with it, mainly log analysis tools, sniffers, and IDS systems.
Plus it's a nice trick to play on script kiddies.
One of the common confusions when setting up your ACLs is why things aren't working as you think they should. For example, you may have given a script access to /etc read/write, and yet it's not functioning properly. Say you have a program /opt/bin/setmotd like the following which changes the /etc/motd file every night:
#!/bin/sh /bin/cp /etc/motd.real /etc/motd /bin/echo >>/etc/motd /bin/uname -a >> /etc/motd /bin/echo "It's `date +%Y/%m/%d`" >>/etc/motd /bin/echo ' -- do you know where your $PPID is?' >> /etc/motd
You've set it up to run via cron, and you give it access to /etc/motd as follows:
lfs# lidsadm -A -s /opt/bin/setmotd -o /etc/motd -j WRITE
However, when it runs, all you see in /etc/motd are the uname and date commands, and you get the following error in your syslogs:
Sep 17 11:20:16 hostname kernel: LIDS: cp (3 2 inode 1037) pid 12680 user (0/0) on pts2: Try to open /etc/motd for writing,flag=33346
Now you've specifically given /opt/bin/setmotd the ability to WRITE to /etc/motd. However the script calls /bin/cp to copy the header file (/etc/motd.real) first, and this is what is failing. The redirect (>>) is done by the shell script itself, and it works as expected.
The problem here is that ACLs are not inherited by their children. When /bin/cp is called, it is the program attempting to rewrite /etc/motd, and it does not have permission. In order to have your /opt/bin/setmotd program work, you need to set the inheritance level of the ACL. (Yes, you could rewrite the script, but there are many cases where you need to set inheritance levels of programs where you'd much rather not rewrite and recompile the software.)
LIDS ACLs can include an inheritance level to allow a subject (program) to inherit the ACL from it's parent. This is done with the '-i' option to lidsadm:
lfs# lidsadm -A -s /opt/bin/setmotd -o /etc/motd -j WRITE -i 1
In the above command we set the inheritance level to 1, meaning that setmotd and its children (in this case /bin/cp) are allowed to WRITE /etc/motd. No grandchildren, however, would inherit this ACL. This is an effective way to pass the use of an ACL from one program to another. You can use an inheritance level of '-1' to denote infinite inheritance. Naturally you should set the inheritance level to the minimum to allow the functionality needed.
Inheritance works for both File and Capability ACLs; however, it is most commonly needed for File ACLs.
Sometimes you may wish you could allow a program to have a special ACL only when it needs it. One of the most common of these situations is when you wish to rotate logs. If you gave the log rotation script the ability to write to the logs directory, an attacker who had gained root access could use it to rotate important logs out of existence to hide her activities. (I personally don't use log rotation scripts that delete logs, I prefer to keep them around forever - hard drive space be damned -- until I manually remove them after archiving.)
So we want to give our log rotation script, /usr/sbin/logrotate, the ability to WRITE to /var/log only when called from cron at the appropriate time. Newer versions of LIDS have the ability to set time restrictions for this ACL, so you'd set up your ACLs as follows:
lfs# crontab -l | grep logrotate 18 0 * * * * /usr/sbin/logrotate # protect /var/log lfs# lidsadm -A -o /var/log -j APPEND # Exception for cron lfs# lidsadm -A -s /usr/sbin/cron -o /var/log \ -t 0018-0019 -i 2 -j WRITE
The above commands give cron WRITE access to /var/log during the window of time when logrotate will be called (00:18) with an inherit level of 2 (to catch logrotate and it's children). This will make sure that logrotate is only able to affect /var/log when called by cron. Had we given logrotate the WRITE access, an attacker could manually run it sufficient times to rotate the logs to /dev/null during that one minute window.
If you are using time-based ACLs, it is important to make sure you crontabs (/var/spool/cron/crontabs, /etc/cron.*) are protected as well, lest an attacker learn what times ACLs may be used to her benefit, or change those crontabs to run her own commands during the vulnerable window.
This concludes the third installment in this series. The next, and final, installment in this series will explore LIDS log entries, using Ptrace, some troublesome issues with LIDS and a brief discussion of what users should be using LIDS to protect.
Brian Hatch is an obsessive security freak and lead author of Hacking Linux Exposed and co-author of Building Linux VPNs. While he frequently stays up late to write or hack code, he thinks it's much more fun to go to the park and push his daughter in the swing as he delivers horrible puns to his fiancee.
This article originally appeared on SecurityFocus.com -- reproduction in whole or in part is not allowed without expressed written consent.